1 // Copyright (C) The Lightning Authors. All rights reserved.
3 // SPDX-License-Identifier: AGPL-3.0
30 "git.arvados.org/arvados.git/sdk/go/arvados"
31 "github.com/arvados/lightning/hgvs"
32 "github.com/james-bowman/nlp"
33 "github.com/kshedden/gonpy"
34 "github.com/sirupsen/logrus"
35 log "github.com/sirupsen/logrus"
36 "golang.org/x/crypto/blake2b"
37 "gonum.org/v1/gonum/mat"
40 const annotationMaxTileSpan = 100
42 type sliceNumpy struct {
47 pvalueMinFrequency float64
57 trainingSet []int // samples index => training set index, or -1 if not in training set
59 pvalue func(onehot []bool) float64
63 func (cmd *sliceNumpy) RunCommand(prog string, args []string, stdin io.Reader, stdout, stderr io.Writer) int {
64 err := cmd.run(prog, args, stdin, stdout, stderr)
66 fmt.Fprintf(stderr, "%s\n", err)
72 func (cmd *sliceNumpy) run(prog string, args []string, stdin io.Reader, stdout, stderr io.Writer) error {
73 flags := flag.NewFlagSet("", flag.ContinueOnError)
74 flags.SetOutput(stderr)
75 pprof := flags.String("pprof", "", "serve Go profile data at http://`[addr]:port`")
76 runlocal := flags.Bool("local", false, "run on local host (default: run in an arvados container)")
77 arvadosRAM := flags.Int("arvados-ram", 750000000000, "amount of memory to request for arvados container (`bytes`)")
78 arvadosVCPUs := flags.Int("arvados-vcpus", 96, "number of VCPUs to request for arvados container")
79 projectUUID := flags.String("project", "", "project `UUID` for output data")
80 priority := flags.Int("priority", 500, "container request priority")
81 preemptible := flags.Bool("preemptible", true, "request preemptible instance")
82 inputDir := flags.String("input-dir", "./in", "input `directory`")
83 outputDir := flags.String("output-dir", "./out", "output `directory`")
84 ref := flags.String("ref", "", "reference name (if blank, choose last one that appears in input)")
85 regionsFilename := flags.String("regions", "", "only output columns/annotations that intersect regions in specified bed `file`")
86 expandRegions := flags.Int("expand-regions", 0, "expand specified regions by `N` base pairs on each side`")
87 mergeOutput := flags.Bool("merge-output", false, "merge output into one matrix.npy and one matrix.annotations.csv")
88 hgvsSingle := flags.Bool("single-hgvs-matrix", false, "also generate hgvs-based matrix")
89 hgvsChunked := flags.Bool("chunked-hgvs-matrix", false, "also generate hgvs-based matrix per chromosome")
90 onehotSingle := flags.Bool("single-onehot", false, "generate one-hot tile-based matrix")
91 onehotChunked := flags.Bool("chunked-onehot", false, "generate one-hot tile-based matrix per input chunk")
92 samplesFilename := flags.String("samples", "", "`samples.csv` file with training/validation and case/control groups (see 'lightning choose-samples')")
93 caseControlOnly := flags.Bool("case-control-only", false, "drop samples that are not in case/control groups")
94 onlyPCA := flags.Bool("pca", false, "run principal component analysis, write components to pca.npy and samples.csv")
95 flags.IntVar(&cmd.pcaComponents, "pca-components", 4, "number of PCA components to compute / use in logistic regression")
96 maxPCATiles := flags.Int("max-pca-tiles", 0, "maximum tiles to use as PCA input (filter, then drop every 2nd colum pair until below max)")
97 debugTag := flags.Int("debug-tag", -1, "log debugging details about specified tag")
98 flags.BoolVar(&cmd.minCoverageAll, "min-coverage-all", false, "apply -min-coverage filter based on all samples, not just training set")
99 flags.IntVar(&cmd.threads, "threads", 16, "number of memory-hungry assembly threads, and number of VCPUs to request for arvados container")
100 flags.Float64Var(&cmd.chi2PValue, "chi2-p-value", 1, "do Χ² test (or logistic regression if -samples file has PCA components) and omit columns with p-value above this threshold")
101 flags.Float64Var(&cmd.pvalueMinFrequency, "pvalue-min-frequency", 0.01, "skip p-value calculation on tile variants below this frequency in the training set")
102 flags.Float64Var(&cmd.maxFrequency, "max-frequency", 1, "do not output variants above this frequency in the training set")
103 flags.BoolVar(&cmd.includeVariant1, "include-variant-1", false, "include most common variant when building one-hot matrix")
104 cmd.filter.Flags(flags)
105 err := flags.Parse(args)
106 if err == flag.ErrHelp {
108 } else if err != nil {
110 } else if flags.NArg() > 0 {
111 return fmt.Errorf("errant command line arguments after parsed flags: %v", flags.Args())
116 log.Println(http.ListenAndServe(*pprof, nil))
120 if cmd.chi2PValue != 1 && *samplesFilename == "" {
121 return fmt.Errorf("cannot use provided -chi2-p-value=%f because -samples= value is empty", cmd.chi2PValue)
124 cmd.debugTag = tagID(*debugTag)
127 runner := arvadosContainerRunner{
128 Name: "lightning slice-numpy",
129 Client: arvados.NewClientFromEnv(),
130 ProjectUUID: *projectUUID,
131 RAM: int64(*arvadosRAM),
132 VCPUs: *arvadosVCPUs,
136 Preemptible: *preemptible,
138 err = runner.TranslatePaths(inputDir, regionsFilename, samplesFilename)
142 runner.Args = []string{"slice-numpy", "-local=true",
144 "-input-dir=" + *inputDir,
145 "-output-dir=/mnt/output",
146 "-threads=" + fmt.Sprintf("%d", cmd.threads),
147 "-regions=" + *regionsFilename,
148 "-expand-regions=" + fmt.Sprintf("%d", *expandRegions),
149 "-merge-output=" + fmt.Sprintf("%v", *mergeOutput),
150 "-single-hgvs-matrix=" + fmt.Sprintf("%v", *hgvsSingle),
151 "-chunked-hgvs-matrix=" + fmt.Sprintf("%v", *hgvsChunked),
152 "-single-onehot=" + fmt.Sprintf("%v", *onehotSingle),
153 "-chunked-onehot=" + fmt.Sprintf("%v", *onehotChunked),
154 "-samples=" + *samplesFilename,
155 "-case-control-only=" + fmt.Sprintf("%v", *caseControlOnly),
156 "-min-coverage-all=" + fmt.Sprintf("%v", cmd.minCoverageAll),
157 "-pca=" + fmt.Sprintf("%v", *onlyPCA),
158 "-pca-components=" + fmt.Sprintf("%d", cmd.pcaComponents),
159 "-max-pca-tiles=" + fmt.Sprintf("%d", *maxPCATiles),
160 "-chi2-p-value=" + fmt.Sprintf("%f", cmd.chi2PValue),
161 "-pvalue-min-frequency=" + fmt.Sprintf("%f", cmd.pvalueMinFrequency),
162 "-max-frequency=" + fmt.Sprintf("%f", cmd.maxFrequency),
163 "-include-variant-1=" + fmt.Sprintf("%v", cmd.includeVariant1),
164 "-debug-tag=" + fmt.Sprintf("%d", cmd.debugTag),
166 runner.Args = append(runner.Args, cmd.filter.Args()...)
168 output, err = runner.Run()
172 fmt.Fprintln(stdout, output)
176 infiles, err := allFiles(*inputDir, matchGobFile)
180 if len(infiles) == 0 {
181 err = fmt.Errorf("no input files found in %s", *inputDir)
184 sort.Strings(infiles)
186 var refseq map[string][]tileLibRef
187 var reftiledata = make(map[tileLibRef][]byte, 11000000)
188 in0, err := open(infiles[0])
193 matchGenome, err := regexp.Compile(cmd.filter.MatchGenome)
195 err = fmt.Errorf("-match-genome: invalid regexp: %q", cmd.filter.MatchGenome)
199 if *samplesFilename != "" {
200 cmd.samples, err = loadSampleInfo(*samplesFilename)
204 } else if *caseControlOnly {
205 return fmt.Errorf("-case-control-only does not make sense without -samples")
210 err = DecodeLibrary(in0, strings.HasSuffix(infiles[0], ".gz"), func(ent *LibraryEntry) error {
211 if len(ent.TagSet) > 0 {
214 for _, cseq := range ent.CompactSequences {
215 if cseq.Name == *ref || *ref == "" {
216 refseq = cseq.TileSequences
219 for _, cg := range ent.CompactGenomes {
220 if matchGenome.MatchString(cg.Name) {
221 cmd.cgnames = append(cmd.cgnames, cg.Name)
224 for _, tv := range ent.TileVariants {
226 reftiledata[tileLibRef{tv.Tag, tv.Variant}] = tv.Sequence
236 err = fmt.Errorf("%s: reference sequence not found", infiles[0])
239 if len(tagset) == 0 {
240 err = fmt.Errorf("tagset not found")
244 taglib := &tagLibrary{}
245 err = taglib.setTags(tagset)
249 taglen := taglib.TagLen()
250 sort.Strings(cmd.cgnames)
252 if len(cmd.cgnames) == 0 {
253 return fmt.Errorf("fatal: 0 matching samples in library, nothing to do")
255 cmd.trainingSet = make([]int, len(cmd.cgnames))
256 if *samplesFilename == "" {
257 cmd.trainingSetSize = len(cmd.cgnames)
258 for i, name := range cmd.cgnames {
259 cmd.samples = append(cmd.samples, sampleInfo{
260 id: trimFilenameForLabel(name),
263 cmd.trainingSet[i] = i
265 } else if len(cmd.cgnames) != len(cmd.samples) {
266 return fmt.Errorf("mismatched sample list: %d samples in library, %d in %s", len(cmd.cgnames), len(cmd.samples), *samplesFilename)
268 for i, name := range cmd.cgnames {
269 if s := trimFilenameForLabel(name); s != cmd.samples[i].id {
270 return fmt.Errorf("mismatched sample list: sample %d is %q in library, %q in %s", i, s, cmd.samples[i].id, *samplesFilename)
273 if *caseControlOnly {
274 for i := 0; i < len(cmd.samples); i++ {
275 if !cmd.samples[i].isTraining && !cmd.samples[i].isValidation {
276 if i+1 < len(cmd.samples) {
277 copy(cmd.samples[i:], cmd.samples[i+1:])
278 copy(cmd.cgnames[i:], cmd.cgnames[i+1:])
280 cmd.samples = cmd.samples[:len(cmd.samples)-1]
281 cmd.cgnames = cmd.cgnames[:len(cmd.cgnames)-1]
287 cmd.trainingSetSize = 0
288 for i := range cmd.cgnames {
289 if cmd.samples[i].isTraining {
290 cmd.trainingSet[i] = cmd.trainingSetSize
291 cmd.trainingSetSize++
292 cmd.chi2Cases = append(cmd.chi2Cases, cmd.samples[i].isCase)
294 cmd.trainingSet[i] = -1
297 if cmd.pvalue == nil {
298 cmd.pvalue = func(onehot []bool) float64 {
299 return pvalue(onehot, cmd.chi2Cases)
304 if cmd.minCoverageAll {
305 cmd.minCoverage = len(cmd.cgnames)
307 cmd.minCoverage = cmd.trainingSetSize
309 if cmd.filter.MinCoverage < 1 {
310 cmd.minCoverage = int(math.Ceil(cmd.filter.MinCoverage * float64(cmd.minCoverage)))
313 if len(cmd.samples[0].pcaComponents) > 0 {
314 cmd.pvalue = glmPvalueFunc(cmd.samples, cmd.pcaComponents)
315 // Unfortunately, statsmodel/glm lib logs stuff to
316 // os.Stdout when it panics on an unsolvable
317 // problem. We recover() from the panic in glm.go, but
318 // we also need to commandeer os.Stdout to avoid
319 // producing large quantities of logs.
320 stdoutWas := os.Stdout
321 defer func() { os.Stdout = stdoutWas }()
322 os.Stdout, err = os.Open(os.DevNull)
328 // cgnamemap[name]==true for samples that we are including in
330 cgnamemap := map[string]bool{}
331 for _, name := range cmd.cgnames {
332 cgnamemap[name] = true
335 err = writeSampleInfo(cmd.samples, *outputDir)
340 log.Info("indexing reference tiles")
341 type reftileinfo struct {
342 variant tileVariantID
343 seqname string // chr1
344 pos int // distance from start of chromosome to starttag
345 tiledata []byte // acgtggcaa...
346 excluded bool // true if excluded by regions file
347 nexttag tagID // tagID of following tile (-1 for last tag of chromosome)
349 isdup := map[tagID]bool{}
350 reftile := map[tagID]*reftileinfo{}
351 for seqname, cseq := range refseq {
353 lastreftag := tagID(-1)
354 for _, libref := range cseq {
355 if cmd.filter.MaxTag >= 0 && libref.Tag > tagID(cmd.filter.MaxTag) {
358 tiledata := reftiledata[libref]
359 if len(tiledata) == 0 {
360 err = fmt.Errorf("missing tiledata for tag %d variant %d in %s in ref", libref.Tag, libref.Variant, seqname)
363 foundthistag := false
364 taglib.FindAll(tiledata[:len(tiledata)-1], func(tagid tagID, offset, _ int) {
365 if !foundthistag && tagid == libref.Tag {
369 if dupref, ok := reftile[tagid]; ok {
370 log.Printf("dropping reference tile %+v from %s @ %d, tag not unique, also found inside %+v from %s @ %d", tileLibRef{Tag: tagid, Variant: dupref.variant}, dupref.seqname, dupref.pos, libref, seqname, pos+offset+1)
371 delete(reftile, tagid)
373 log.Printf("found tag %d at offset %d inside tile variant %+v on %s @ %d", tagid, offset, libref, seqname, pos+offset+1)
377 if isdup[libref.Tag] {
378 log.Printf("dropping reference tile %+v from %s @ %d, tag not unique", libref, seqname, pos)
379 } else if reftile[libref.Tag] != nil {
380 log.Printf("dropping reference tile %+v from %s @ %d, tag not unique", tileLibRef{Tag: libref.Tag, Variant: reftile[libref.Tag].variant}, reftile[libref.Tag].seqname, reftile[libref.Tag].pos)
381 delete(reftile, libref.Tag)
382 log.Printf("dropping reference tile %+v from %s @ %d, tag not unique", libref, seqname, pos)
383 isdup[libref.Tag] = true
385 reftile[libref.Tag] = &reftileinfo{
387 variant: libref.Variant,
393 reftile[lastreftag].nexttag = libref.Tag
395 lastreftag = libref.Tag
397 pos += len(tiledata) - taglen
399 log.Printf("... %s done, len %d", seqname, pos+taglen)
403 if *regionsFilename != "" {
404 log.Printf("loading regions from %s", *regionsFilename)
405 mask, err = makeMask(*regionsFilename, *expandRegions)
409 log.Printf("before applying mask, len(reftile) == %d", len(reftile))
410 log.Printf("deleting reftile entries for regions outside %d intervals", mask.Len())
411 for _, rt := range reftile {
412 if !mask.Check(strings.TrimPrefix(rt.seqname, "chr"), rt.pos, rt.pos+len(rt.tiledata)) {
416 log.Printf("after applying mask, len(reftile) == %d", len(reftile))
419 type hgvsColSet map[hgvs.Variant][2][]int8
420 encodeHGVS := throttle{Max: len(refseq)}
421 encodeHGVSTodo := map[string]chan hgvsColSet{}
422 tmpHGVSCols := map[string]*os.File{}
424 for seqname := range refseq {
426 f, err = os.Create(*outputDir + "/tmp." + seqname + ".gob")
430 defer os.Remove(f.Name())
431 bufw := bufio.NewWriterSize(f, 1<<24)
432 enc := gob.NewEncoder(bufw)
433 tmpHGVSCols[seqname] = f
434 todo := make(chan hgvsColSet, 128)
435 encodeHGVSTodo[seqname] = todo
436 encodeHGVS.Go(func() error {
437 for colset := range todo {
438 err := enc.Encode(colset)
440 encodeHGVS.Report(err)
451 var toMerge [][]int16
452 if *mergeOutput || *hgvsSingle {
453 toMerge = make([][]int16, len(infiles))
455 var onehotIndirect [][2][]uint32 // [chunkIndex][axis][index]
456 var onehotChunkSize []uint32
457 var onehotXrefs [][]onehotXref
458 if *onehotSingle || *onlyPCA {
459 onehotIndirect = make([][2][]uint32, len(infiles))
460 onehotChunkSize = make([]uint32, len(infiles))
461 onehotXrefs = make([][]onehotXref, len(infiles))
463 chunkStartTag := make([]tagID, len(infiles))
465 throttleMem := throttle{Max: cmd.threads} // TODO: estimate using mem and data size
466 throttleNumpyMem := throttle{Max: cmd.threads/2 + 1}
467 log.Info("generating annotations and numpy matrix for each slice")
468 var errSkip = errors.New("skip infile")
470 for infileIdx, infile := range infiles {
471 infileIdx, infile := infileIdx, infile
472 throttleMem.Go(func() error {
473 seq := make(map[tagID][]TileVariant, 50000)
474 cgs := make(map[string]CompactGenome, len(cmd.cgnames))
475 f, err := open(infile)
480 log.Infof("%04d: reading %s", infileIdx, infile)
481 err = DecodeLibrary(f, strings.HasSuffix(infile, ".gz"), func(ent *LibraryEntry) error {
482 for _, tv := range ent.TileVariants {
487 // corresponding ref tile, if
488 // mask is in play (we can't
489 // determine coordinates for
491 if mask != nil && reftile[tv.Tag] == nil {
495 // corresponding ref tile is
496 // outside target regions --
497 // unless it's a potential
499 if mask != nil && reftile[tv.Tag].excluded &&
500 (int(tv.Tag+1) >= len(tagset) ||
501 (bytes.HasSuffix(tv.Sequence, tagset[tv.Tag+1]) && reftile[tv.Tag+1] != nil && !reftile[tv.Tag+1].excluded)) {
504 if tv.Tag == cmd.debugTag {
505 log.Printf("infile %d %s tag %d variant %d hash %x", infileIdx, infile, tv.Tag, tv.Variant, tv.Blake2b[:3])
507 variants := seq[tv.Tag]
508 if len(variants) == 0 {
509 variants = make([]TileVariant, 100)
511 for len(variants) <= int(tv.Variant) {
512 variants = append(variants, TileVariant{})
514 variants[int(tv.Variant)] = tv
515 seq[tv.Tag] = variants
517 for _, cg := range ent.CompactGenomes {
518 if cmd.filter.MaxTag >= 0 && cg.StartTag > tagID(cmd.filter.MaxTag) {
521 if !cgnamemap[cg.Name] {
524 // pad to full slice size
525 // to avoid out-of-bounds
527 if sliceSize := 2 * int(cg.EndTag-cg.StartTag); len(cg.Variants) < sliceSize {
528 cg.Variants = append(cg.Variants, make([]tileVariantID, sliceSize-len(cg.Variants))...)
536 } else if err != nil {
537 return fmt.Errorf("%04d: DecodeLibrary(%s): %w", infileIdx, infile, err)
539 tagstart := cgs[cmd.cgnames[0]].StartTag
540 tagend := cgs[cmd.cgnames[0]].EndTag
541 chunkStartTag[infileIdx] = tagstart
545 log.Infof("%04d: renumber/dedup variants for tags %d-%d", infileIdx, tagstart, tagend)
546 variantRemap := make([][]tileVariantID, tagend-tagstart)
547 throttleCPU := throttle{Max: runtime.GOMAXPROCS(0)}
548 for tag, variants := range seq {
549 tag, variants := tag, variants
550 throttleCPU.Go(func() error {
552 count := make(map[[blake2b.Size256]byte]int, len(variants))
556 count[blake2b.Sum256(rt.tiledata)] = 0
559 for cgidx, cgname := range cmd.cgnames {
560 if !cmd.minCoverageAll && !cmd.samples[cgidx].isTraining {
564 idx := int(tag-tagstart) * 2
565 for allele := 0; allele < 2; allele++ {
566 v := cg.Variants[idx+allele]
567 if v > 0 && len(variants[v].Sequence) > 0 {
568 count[variants[v].Blake2b]++
571 if v > 0 && tag == cmd.debugTag {
572 log.Printf("tag %d cg %s allele %d tv %d hash %x count is now %d", tag, cgname, allele, v, variants[v].Blake2b[:3], count[variants[v].Blake2b])
576 if alleleCoverage < cmd.minCoverage*2 {
577 idx := int(tag-tagstart) * 2
578 for _, cg := range cgs {
580 cg.Variants[idx+1] = 0
582 if tag == cmd.debugTag {
583 log.Printf("tag %d alleleCoverage %d < min %d, sample data wiped", tag, alleleCoverage, cmd.minCoverage*2)
588 // hash[i] will be the hash of
589 // the variant(s) that should
590 // be at rank i (0-based).
591 hash := make([][blake2b.Size256]byte, 0, len(count))
592 for b := range count {
593 hash = append(hash, b)
595 sort.Slice(hash, func(i, j int) bool {
596 bi, bj := &hash[i], &hash[j]
597 if ci, cj := count[*bi], count[*bj]; ci != cj {
600 return bytes.Compare((*bi)[:], (*bj)[:]) < 0
603 // rank[b] will be the 1-based
604 // new variant number for
605 // variants whose hash is b.
606 rank := make(map[[blake2b.Size256]byte]tileVariantID, len(hash))
607 for i, h := range hash {
608 rank[h] = tileVariantID(i + 1)
610 if tag == cmd.debugTag {
611 for h, r := range rank {
612 log.Printf("tag %d rank(%x) = %v", tag, h[:3], r)
615 // remap[v] will be the new
616 // variant number for original
618 remap := make([]tileVariantID, len(variants))
619 for i, tv := range variants {
620 remap[i] = rank[tv.Blake2b]
622 if tag == cmd.debugTag {
623 for in, out := range remap {
625 log.Printf("tag %d remap %d => %d", tag, in, out)
629 variantRemap[tag-tagstart] = remap
631 refrank := rank[blake2b.Sum256(rt.tiledata)]
632 if tag == cmd.debugTag {
633 log.Printf("tag %d reftile variant %d => %d", tag, rt.variant, refrank)
642 var onehotChunk [][]int8
643 var onehotXref []onehotXref
645 var annotationsFilename string
647 annotationsFilename = "/dev/null"
649 annotationsFilename = fmt.Sprintf("%s/matrix.%04d.annotations.csv", *outputDir, infileIdx)
650 log.Infof("%04d: writing %s", infileIdx, annotationsFilename)
652 annof, err := os.Create(annotationsFilename)
656 annow := bufio.NewWriterSize(annof, 1<<20)
658 for tag := tagstart; tag < tagend; tag++ {
660 if rt == nil && mask != nil {
661 // With no ref tile, we don't
662 // have coordinates to say
663 // this is in the desired
664 // regions -- so it's not.
665 // TODO: handle ref spanning
669 if rt != nil && rt.excluded {
670 // TODO: don't skip yet --
671 // first check for spanning
672 // tile variants that
673 // intersect non-excluded ref
677 if cmd.filter.MaxTag >= 0 && tag > tagID(cmd.filter.MaxTag) {
680 remap := variantRemap[tag-tagstart]
682 // was not assigned above,
683 // because minCoverage
687 maxv := tileVariantID(0)
688 for _, v := range remap {
693 if *onehotChunked || *onehotSingle || *onlyPCA {
694 onehot, xrefs := cmd.tv2homhet(cgs, maxv, remap, tag, tagstart, seq)
695 if tag == cmd.debugTag {
696 log.WithFields(logrus.Fields{
699 }).Info("tv2homhet()")
701 onehotChunk = append(onehotChunk, onehot...)
702 onehotXref = append(onehotXref, xrefs...)
709 // Reference does not use any
710 // variant of this tile
712 // TODO: diff against the
713 // relevant portion of the
714 // ref's spanning tile
718 fmt.Fprintf(annow, "%d,%d,%d,=,%s,%d,,,\n", tag, outcol, rt.variant, rt.seqname, rt.pos)
720 reftilestr := strings.ToUpper(string(rt.tiledata))
722 done := make([]bool, maxv+1)
723 variantDiffs := make([][]hgvs.Variant, maxv+1)
724 for v, tv := range variants {
726 if v == 0 || v == rt.variant || done[v] {
731 if len(tv.Sequence) < taglen {
734 // if reftilestr doesn't end
735 // in the same tag as tv,
736 // extend reftilestr with
737 // following ref tiles until
738 // it does (up to an arbitrary
739 // sanity-check limit)
740 reftilestr := reftilestr
741 endtagstr := strings.ToUpper(string(tv.Sequence[len(tv.Sequence)-taglen:]))
742 for i, rt := 0, rt; i < annotationMaxTileSpan && !strings.HasSuffix(reftilestr, endtagstr) && rt.nexttag >= 0; i++ {
743 rt = reftile[rt.nexttag]
747 reftilestr += strings.ToUpper(string(rt.tiledata[taglen:]))
749 if mask != nil && !mask.Check(strings.TrimPrefix(rt.seqname, "chr"), rt.pos, rt.pos+len(reftilestr)) {
752 if !strings.HasSuffix(reftilestr, endtagstr) {
753 fmt.Fprintf(annow, "%d,%d,%d,,%s,%d,,,\n", tag, outcol, v, rt.seqname, rt.pos)
756 if lendiff := len(reftilestr) - len(tv.Sequence); lendiff < -1000 || lendiff > 1000 {
757 fmt.Fprintf(annow, "%d,%d,%d,,%s,%d,,,\n", tag, outcol, v, rt.seqname, rt.pos)
760 diffs, _ := hgvs.Diff(reftilestr, strings.ToUpper(string(tv.Sequence)), 0)
761 for i := range diffs {
762 diffs[i].Position += rt.pos
764 for _, diff := range diffs {
765 fmt.Fprintf(annow, "%d,%d,%d,%s:g.%s,%s,%d,%s,%s,%s\n", tag, outcol, v, rt.seqname, diff.String(), rt.seqname, diff.Position, diff.Ref, diff.New, diff.Left)
768 variantDiffs[v] = diffs
772 // We can now determine, for each HGVS
773 // variant (diff) in this reftile
774 // region, whether a given genome
775 // phase/allele (1) has the variant, (0) has
776 // =ref or a different variant in that
777 // position, or (-1) is lacking
778 // coverage / couldn't be diffed.
779 hgvsCol := hgvsColSet{}
780 for _, diffs := range variantDiffs {
781 for _, diff := range diffs {
782 if _, ok := hgvsCol[diff]; ok {
785 hgvsCol[diff] = [2][]int8{
786 make([]int8, len(cmd.cgnames)),
787 make([]int8, len(cmd.cgnames)),
791 for row, name := range cmd.cgnames {
792 variants := cgs[name].Variants[(tag-tagstart)*2:]
793 for ph := 0; ph < 2; ph++ {
795 if int(v) >= len(remap) {
801 // hgvsCol[*][ph][row] is already 0
802 } else if len(variantDiffs[v]) == 0 {
803 // lacking coverage / couldn't be diffed
804 for _, col := range hgvsCol {
808 for _, diff := range variantDiffs[v] {
809 hgvsCol[diff][ph][row] = 1
814 for diff, colpair := range hgvsCol {
815 allele2homhet(colpair)
816 if !cmd.filterHGVScolpair(colpair) {
817 delete(hgvsCol, diff)
820 if len(hgvsCol) > 0 {
821 encodeHGVSTodo[rt.seqname] <- hgvsCol
836 // transpose onehotChunk[col][row] to numpy[row*ncols+col]
837 rows := len(cmd.cgnames)
838 cols := len(onehotChunk)
839 log.Infof("%04d: preparing onehot numpy (rows=%d, cols=%d, mem=%d)", infileIdx, rows, cols, rows*cols)
840 throttleNumpyMem.Acquire()
841 out := onehotcols2int8(onehotChunk)
842 fnm := fmt.Sprintf("%s/onehot.%04d.npy", *outputDir, infileIdx)
843 err = writeNumpyInt8(fnm, out, rows, cols)
847 fnm = fmt.Sprintf("%s/onehot-columns.%04d.npy", *outputDir, infileIdx)
848 err = writeNumpyInt32(fnm, onehotXref2int32(onehotXref), 4, len(onehotXref))
853 throttleNumpyMem.Release()
855 if *onehotSingle || *onlyPCA {
856 onehotIndirect[infileIdx] = onehotChunk2Indirect(onehotChunk)
857 onehotChunkSize[infileIdx] = uint32(len(onehotChunk))
858 onehotXrefs[infileIdx] = onehotXref
859 n := len(onehotIndirect[infileIdx][0])
860 log.Infof("%04d: keeping onehot coordinates in memory (n=%d, mem=%d)", infileIdx, n, n*8*2)
862 if !(*onehotSingle || *onehotChunked || *onlyPCA) || *mergeOutput || *hgvsSingle {
863 log.Infof("%04d: preparing numpy (rows=%d, cols=%d)", infileIdx, len(cmd.cgnames), 2*outcol)
864 throttleNumpyMem.Acquire()
865 rows := len(cmd.cgnames)
867 out := make([]int16, rows*cols)
868 for row, name := range cmd.cgnames {
870 for col, v := range cgs[name].Variants {
871 tag := tagstart + tagID(col/2)
872 if cmd.filter.MaxTag >= 0 && tag > tagID(cmd.filter.MaxTag) {
875 if rt := reftile[tag]; mask != nil && (rt == nil || rt.excluded) {
879 out[outidx] = 0 // tag not found / spanning tile
880 } else if variants, ok := seq[tag]; ok && int(v) < len(variants) && len(variants[v].Sequence) > 0 {
881 out[outidx] = int16(variantRemap[tag-tagstart][v])
883 out[outidx] = -1 // low quality tile variant
885 if tag == cmd.debugTag {
886 log.Printf("tag %d row %d col %d outidx %d v %d out %d", tag, row, col, outidx, v, out[outidx])
894 throttleNumpyMem.Release()
895 if *mergeOutput || *hgvsSingle {
896 log.Infof("%04d: matrix fragment %d rows x %d cols", infileIdx, rows, cols)
897 toMerge[infileIdx] = out
899 if !*mergeOutput && !*onehotChunked && !*onehotSingle {
900 fnm := fmt.Sprintf("%s/matrix.%04d.npy", *outputDir, infileIdx)
901 err = writeNumpyInt16(fnm, out, rows, cols)
908 log.Infof("%s: done (%d/%d)", infile, int(atomic.AddInt64(&done, 1)), len(infiles))
912 if err = throttleMem.Wait(); err != nil {
917 log.Info("flushing hgvsCols temp files")
918 for seqname := range refseq {
919 close(encodeHGVSTodo[seqname])
921 err = encodeHGVS.Wait()
925 for seqname := range refseq {
926 log.Infof("%s: reading hgvsCols from temp file", seqname)
927 f := tmpHGVSCols[seqname]
928 _, err = f.Seek(0, io.SeekStart)
932 var hgvsCols hgvsColSet
933 dec := gob.NewDecoder(bufio.NewReaderSize(f, 1<<24))
935 err = dec.Decode(&hgvsCols)
940 log.Infof("%s: sorting %d hgvs variants", seqname, len(hgvsCols))
941 variants := make([]hgvs.Variant, 0, len(hgvsCols))
942 for v := range hgvsCols {
943 variants = append(variants, v)
945 sort.Slice(variants, func(i, j int) bool {
946 vi, vj := &variants[i], &variants[j]
947 if vi.Position != vj.Position {
948 return vi.Position < vj.Position
949 } else if vi.Ref != vj.Ref {
950 return vi.Ref < vj.Ref
952 return vi.New < vj.New
955 rows := len(cmd.cgnames)
956 cols := len(variants) * 2
957 log.Infof("%s: building hgvs matrix (rows=%d, cols=%d, mem=%d)", seqname, rows, cols, rows*cols)
958 out := make([]int8, rows*cols)
959 for varIdx, variant := range variants {
960 hgvsCols := hgvsCols[variant]
961 for row := range cmd.cgnames {
962 for ph := 0; ph < 2; ph++ {
963 out[row*cols+varIdx+ph] = hgvsCols[ph][row]
967 err = writeNumpyInt8(fmt.Sprintf("%s/hgvs.%s.npy", *outputDir, seqname), out, rows, cols)
973 fnm := fmt.Sprintf("%s/hgvs.%s.annotations.csv", *outputDir, seqname)
974 log.Infof("%s: writing hgvs column labels to %s", seqname, fnm)
975 var hgvsLabels bytes.Buffer
976 for varIdx, variant := range variants {
977 fmt.Fprintf(&hgvsLabels, "%d,%s:g.%s\n", varIdx, seqname, variant.String())
979 err = ioutil.WriteFile(fnm, hgvsLabels.Bytes(), 0666)
986 if *mergeOutput || *hgvsSingle {
987 var annow *bufio.Writer
990 annoFilename := fmt.Sprintf("%s/matrix.annotations.csv", *outputDir)
991 annof, err = os.Create(annoFilename)
995 annow = bufio.NewWriterSize(annof, 1<<20)
998 rows := len(cmd.cgnames)
1000 for _, chunk := range toMerge {
1001 cols += len(chunk) / rows
1003 log.Infof("merging output matrix (rows=%d, cols=%d, mem=%d) and annotations", rows, cols, rows*cols*2)
1006 out = make([]int16, rows*cols)
1008 hgvsCols := map[string][2][]int16{} // hgvs -> [[g0,g1,g2,...], [g0,g1,g2,...]] (slice of genomes for each phase)
1010 for outIdx, chunk := range toMerge {
1011 chunkcols := len(chunk) / rows
1013 for row := 0; row < rows; row++ {
1014 copy(out[row*cols+startcol:], chunk[row*chunkcols:(row+1)*chunkcols])
1017 toMerge[outIdx] = nil
1019 annotationsFilename := fmt.Sprintf("%s/matrix.%04d.annotations.csv", *outputDir, outIdx)
1020 log.Infof("reading %s", annotationsFilename)
1021 buf, err := os.ReadFile(annotationsFilename)
1026 err = os.Remove(annotationsFilename)
1031 for _, line := range bytes.Split(buf, []byte{'\n'}) {
1035 fields := bytes.SplitN(line, []byte{','}, 9)
1036 tag, _ := strconv.Atoi(string(fields[0]))
1037 incol, _ := strconv.Atoi(string(fields[1]))
1038 tileVariant, _ := strconv.Atoi(string(fields[2]))
1039 hgvsID := string(fields[3])
1040 seqname := string(fields[4])
1041 pos, _ := strconv.Atoi(string(fields[5]))
1044 // Null entry for un-diffable
1049 // Null entry for ref tile
1052 if mask != nil && !mask.Check(strings.TrimPrefix(seqname, "chr"), pos, pos+len(refseq)) {
1053 // The tile intersects one of
1054 // the selected regions, but
1055 // this particular HGVS
1056 // variant does not.
1059 hgvsColPair := hgvsCols[hgvsID]
1060 if hgvsColPair[0] == nil {
1061 // values in new columns start
1062 // out as -1 ("no data yet")
1063 // or 0 ("=ref") here, may
1064 // change to 1 ("hgvs variant
1065 // present") below, either on
1066 // this line or a future line.
1067 hgvsColPair = [2][]int16{make([]int16, len(cmd.cgnames)), make([]int16, len(cmd.cgnames))}
1068 rt, ok := reftile[tagID(tag)]
1070 err = fmt.Errorf("bug: seeing annotations for tag %d, but it has no reftile entry", tag)
1073 for ph := 0; ph < 2; ph++ {
1074 for row := 0; row < rows; row++ {
1075 v := chunk[row*chunkcols+incol*2+ph]
1076 if tileVariantID(v) == rt.variant {
1077 hgvsColPair[ph][row] = 0
1079 hgvsColPair[ph][row] = -1
1083 hgvsCols[hgvsID] = hgvsColPair
1085 hgvsref := hgvs.Variant{
1087 Ref: string(refseq),
1088 New: string(refseq),
1090 fmt.Fprintf(annow, "%d,%d,%d,%s:g.%s,%s,%d,%s,%s,%s\n", tag, incol+startcol/2, rt.variant, seqname, hgvsref.String(), seqname, pos, refseq, refseq, fields[8])
1094 fmt.Fprintf(annow, "%d,%d,%d,%s,%s,%d,%s,%s,%s\n", tag, incol+startcol/2, tileVariant, hgvsID, seqname, pos, refseq, fields[7], fields[8])
1096 for ph := 0; ph < 2; ph++ {
1097 for row := 0; row < rows; row++ {
1098 v := chunk[row*chunkcols+incol*2+ph]
1099 if int(v) == tileVariant {
1100 hgvsColPair[ph][row] = 1
1106 startcol += chunkcols
1117 err = writeNumpyInt16(fmt.Sprintf("%s/matrix.npy", *outputDir), out, rows, cols)
1125 cols = len(hgvsCols) * 2
1126 log.Printf("building hgvs-based matrix: %d rows x %d cols", rows, cols)
1127 out = make([]int16, rows*cols)
1128 hgvsIDs := make([]string, 0, cols/2)
1129 for hgvsID := range hgvsCols {
1130 hgvsIDs = append(hgvsIDs, hgvsID)
1132 sort.Strings(hgvsIDs)
1133 var hgvsLabels bytes.Buffer
1134 for idx, hgvsID := range hgvsIDs {
1135 fmt.Fprintf(&hgvsLabels, "%d,%s\n", idx, hgvsID)
1136 for ph := 0; ph < 2; ph++ {
1137 hgvscol := hgvsCols[hgvsID][ph]
1138 for row, val := range hgvscol {
1139 out[row*cols+idx*2+ph] = val
1143 err = writeNumpyInt16(fmt.Sprintf("%s/hgvs.npy", *outputDir), out, rows, cols)
1148 fnm := fmt.Sprintf("%s/hgvs.annotations.csv", *outputDir)
1149 log.Printf("writing hgvs labels: %s", fnm)
1150 err = ioutil.WriteFile(fnm, hgvsLabels.Bytes(), 0777)
1156 if *onehotSingle || *onlyPCA {
1158 for _, part := range onehotIndirect {
1159 nzCount += len(part[0])
1161 onehot := make([]uint32, nzCount*2) // [r,r,r,...,c,c,c,...]
1162 var xrefs []onehotXref
1163 chunkOffset := uint32(0)
1165 for i, part := range onehotIndirect {
1166 for i := range part[1] {
1167 part[1][i] += chunkOffset
1169 copy(onehot[outcol:], part[0])
1170 copy(onehot[outcol+nzCount:], part[1])
1171 xrefs = append(xrefs, onehotXrefs[i]...)
1173 outcol += len(part[0])
1174 chunkOffset += onehotChunkSize[i]
1178 onehotXrefs[i] = nil
1179 debug.FreeOSMemory()
1182 fnm := fmt.Sprintf("%s/onehot.npy", *outputDir)
1183 err = writeNumpyUint32(fnm, onehot, 2, nzCount)
1187 fnm = fmt.Sprintf("%s/onehot-columns.npy", *outputDir)
1188 err = writeNumpyInt32(fnm, onehotXref2int32(xrefs), 5, len(xrefs))
1192 fnm = fmt.Sprintf("%s/stats.json", *outputDir)
1193 j, err := json.Marshal(map[string]interface{}{
1194 "pvalueCallCount": cmd.pvalueCallCount,
1199 err = os.WriteFile(fnm, j, 0777)
1206 for _, c := range onehot[nzCount:] {
1212 return fmt.Errorf("cannot do PCA: one-hot matrix is empty")
1214 log.Printf("have %d one-hot cols", cols)
1216 for *maxPCATiles > 0 && cols > *maxPCATiles*2 {
1217 cols = (cols + 1) / 2
1221 // we work with pairs of columns
1224 log.Printf("creating full matrix (%d rows) and training matrix (%d rows) with %d cols, stride %d", len(cmd.cgnames), cmd.trainingSetSize, cols, stride)
1225 mtxFull := mat.NewDense(len(cmd.cgnames), cols, nil)
1226 mtxTrain := mat.NewDense(cmd.trainingSetSize, cols, nil)
1227 for i, c := range onehot[nzCount:] {
1228 if int(c/2)%stride == 0 {
1229 outcol := int(c/2)/stride*2 + int(c)%2
1230 mtxFull.Set(int(onehot[i]), outcol, 1)
1231 if trainRow := cmd.trainingSet[int(onehot[i])]; trainRow >= 0 {
1232 mtxTrain.Set(trainRow, outcol, 1)
1236 log.Print("fitting")
1237 transformer := nlp.NewPCA(cmd.pcaComponents)
1238 transformer.Fit(mtxTrain.T())
1239 log.Printf("transforming")
1240 pca, err := transformer.Transform(mtxFull.T())
1245 outrows, outcols := pca.Dims()
1246 log.Printf("copying result to numpy output array: %d rows, %d cols", outrows, outcols)
1247 out := make([]float64, outrows*outcols)
1248 for i := 0; i < outrows; i++ {
1249 for j := 0; j < outcols; j++ {
1250 out[i*outcols+j] = pca.At(i, j)
1253 fnm := fmt.Sprintf("%s/pca.npy", *outputDir)
1254 log.Printf("writing numpy: %s", fnm)
1255 output, err := os.OpenFile(fnm, os.O_CREATE|os.O_TRUNC|os.O_WRONLY, 0777)
1259 npw, err := gonpy.NewWriter(nopCloser{output})
1261 return fmt.Errorf("gonpy.NewWriter: %w", err)
1263 npw.Shape = []int{outrows, outcols}
1264 err = npw.WriteFloat64(out)
1266 return fmt.Errorf("WriteFloat64: %w", err)
1268 err = output.Close()
1274 log.Print("copying pca components to sampleInfo")
1275 for i := range cmd.samples {
1276 cmd.samples[i].pcaComponents = make([]float64, outcols)
1277 for c := 0; c < outcols; c++ {
1278 cmd.samples[i].pcaComponents[c] = pca.At(i, c)
1283 err = writeSampleInfo(cmd.samples, *outputDir)
1289 if !*mergeOutput && !*onehotChunked && !*onehotSingle && !*onlyPCA {
1290 tagoffsetFilename := *outputDir + "/chunk-tag-offset.csv"
1291 log.Infof("writing tag offsets to %s", tagoffsetFilename)
1293 f, err = os.Create(tagoffsetFilename)
1298 for idx, offset := range chunkStartTag {
1299 _, err = fmt.Fprintf(f, "%q,%d\n", fmt.Sprintf("matrix.%04d.npy", idx), offset)
1301 err = fmt.Errorf("write %s: %w", tagoffsetFilename, err)
1307 err = fmt.Errorf("close %s: %w", tagoffsetFilename, err)
1315 type sampleInfo struct {
1321 pcaComponents []float64
1324 // Read samples.csv file with case/control and training/validation
1326 func loadSampleInfo(samplesFilename string) ([]sampleInfo, error) {
1328 f, err := open(samplesFilename)
1332 buf, err := io.ReadAll(f)
1338 for _, csv := range bytes.Split(buf, []byte{'\n'}) {
1343 split := strings.Split(string(csv), ",")
1345 return nil, fmt.Errorf("%d fields < 4 in %s line %d: %q", len(split), samplesFilename, lineNum, csv)
1347 if split[0] == "Index" && split[1] == "SampleID" && split[2] == "CaseControl" && split[3] == "TrainingValidation" {
1350 idx, err := strconv.Atoi(split[0])
1353 return nil, fmt.Errorf("header does not look right: %q", csv)
1355 return nil, fmt.Errorf("%s line %d: index: %s", samplesFilename, lineNum, err)
1358 return nil, fmt.Errorf("%s line %d: index %d out of order", samplesFilename, lineNum, idx)
1360 var pcaComponents []float64
1362 for _, s := range split[4:] {
1363 f, err := strconv.ParseFloat(s, 64)
1365 return nil, fmt.Errorf("%s line %d: cannot parse float %q: %s", samplesFilename, lineNum, s, err)
1367 pcaComponents = append(pcaComponents, f)
1370 si = append(si, sampleInfo{
1372 isCase: split[2] == "1",
1373 isControl: split[2] == "0",
1374 isTraining: split[3] == "1",
1375 isValidation: split[3] == "0" && len(split[2]) > 0, // fix errant 0s in input
1376 pcaComponents: pcaComponents,
1382 func writeSampleInfo(samples []sampleInfo, outputDir string) error {
1383 fnm := outputDir + "/samples.csv"
1384 log.Infof("writing sample metadata to %s", fnm)
1385 f, err := os.Create(fnm)
1391 if len(samples) > 0 {
1392 for i := range samples[0].pcaComponents {
1393 pcaLabels += fmt.Sprintf(",PCA%d", i)
1396 _, err = fmt.Fprintf(f, "Index,SampleID,CaseControl,TrainingValidation%s\n", pcaLabels)
1400 for i, si := range samples {
1404 } else if si.isControl {
1409 } else if si.isValidation {
1413 for _, pcaval := range si.pcaComponents {
1414 pcavals += fmt.Sprintf(",%f", pcaval)
1416 _, err = fmt.Fprintf(f, "%d,%s,%s,%s%s\n", i, si.id, cc, tv, pcavals)
1418 return fmt.Errorf("write %s: %w", fnm, err)
1423 return fmt.Errorf("close %s: %w", fnm, err)
1429 func (cmd *sliceNumpy) filterHGVScolpair(colpair [2][]int8) bool {
1430 if cmd.chi2PValue >= 1 {
1433 col0 := make([]bool, 0, len(cmd.chi2Cases))
1434 col1 := make([]bool, 0, len(cmd.chi2Cases))
1435 cases := make([]bool, 0, len(cmd.chi2Cases))
1436 for i, c := range cmd.chi2Cases {
1437 if colpair[0][i] < 0 {
1440 col0 = append(col0, colpair[0][i] != 0)
1441 col1 = append(col1, colpair[1][i] != 0)
1442 cases = append(cases, c)
1444 return len(cases) >= cmd.minCoverage &&
1445 (pvalue(col0, cases) <= cmd.chi2PValue || pvalue(col1, cases) <= cmd.chi2PValue)
1448 func writeNumpyUint32(fnm string, out []uint32, rows, cols int) error {
1449 output, err := os.Create(fnm)
1453 defer output.Close()
1454 bufw := bufio.NewWriterSize(output, 1<<26)
1455 npw, err := gonpy.NewWriter(nopCloser{bufw})
1459 log.WithFields(log.Fields{
1463 "bytes": rows * cols * 4,
1464 }).Infof("writing numpy: %s", fnm)
1465 npw.Shape = []int{rows, cols}
1466 npw.WriteUint32(out)
1471 return output.Close()
1474 func writeNumpyInt32(fnm string, out []int32, rows, cols int) error {
1475 output, err := os.Create(fnm)
1479 defer output.Close()
1480 bufw := bufio.NewWriterSize(output, 1<<26)
1481 npw, err := gonpy.NewWriter(nopCloser{bufw})
1485 log.WithFields(log.Fields{
1489 "bytes": rows * cols * 4,
1490 }).Infof("writing numpy: %s", fnm)
1491 npw.Shape = []int{rows, cols}
1497 return output.Close()
1500 func writeNumpyInt16(fnm string, out []int16, rows, cols int) error {
1501 output, err := os.Create(fnm)
1505 defer output.Close()
1506 bufw := bufio.NewWriterSize(output, 1<<26)
1507 npw, err := gonpy.NewWriter(nopCloser{bufw})
1511 log.WithFields(log.Fields{
1515 "bytes": rows * cols * 2,
1516 }).Infof("writing numpy: %s", fnm)
1517 npw.Shape = []int{rows, cols}
1523 return output.Close()
1526 func writeNumpyInt8(fnm string, out []int8, rows, cols int) error {
1527 output, err := os.Create(fnm)
1531 defer output.Close()
1532 bufw := bufio.NewWriterSize(output, 1<<26)
1533 npw, err := gonpy.NewWriter(nopCloser{bufw})
1537 log.WithFields(log.Fields{
1541 "bytes": rows * cols,
1542 }).Infof("writing numpy: %s", fnm)
1543 npw.Shape = []int{rows, cols}
1549 return output.Close()
1552 func allele2homhet(colpair [2][]int8) {
1553 a, b := colpair[0], colpair[1]
1554 for i, av := range a {
1556 if av < 0 || bv < 0 {
1559 } else if av > 0 && bv > 0 {
1562 } else if av > 0 || bv > 0 {
1566 // ref (or a different variant in same position)
1567 // (this is a no-op) a[i], b[i] = 0, 0
1572 type onehotXref struct {
1574 variant tileVariantID
1580 const onehotXrefSize = unsafe.Sizeof(onehotXref{})
1582 // Build onehot matrix (m[tileVariantIndex][genome] == 0 or 1) for all
1583 // variants of a single tile/tag#.
1585 // Return nil if no tile variant passes Χ² filter.
1586 func (cmd *sliceNumpy) tv2homhet(cgs map[string]CompactGenome, maxv tileVariantID, remap []tileVariantID, tag, chunkstarttag tagID, seq map[tagID][]TileVariant) ([][]int8, []onehotXref) {
1587 if tag == cmd.debugTag {
1588 tv := make([]tileVariantID, len(cmd.cgnames)*2)
1589 for i, name := range cmd.cgnames {
1590 copy(tv[i*2:(i+1)*2], cgs[name].Variants[(tag-chunkstarttag)*2:])
1592 log.WithFields(logrus.Fields{
1593 "cgs[i].Variants[tag*2+j]": tv,
1597 "chunkstarttag": chunkstarttag,
1598 }).Info("tv2homhet()")
1600 if maxv < 1 || (maxv < 2 && !cmd.includeVariant1) {
1601 // everyone has the most common variant (of the variants we don't drop)
1604 tagoffset := tag - chunkstarttag
1606 for cgidx, cgname := range cmd.cgnames {
1607 if !cmd.minCoverageAll && !cmd.samples[cgidx].isTraining {
1612 for _, v := range cg.Variants[tagoffset*2 : tagoffset*2+2] {
1613 if v > 0 && int(v) < len(seq[tag]) && len(seq[tag][v].Sequence) > 0 {
1621 if coverage < cmd.minCoverage {
1624 // "observed" array for p-value calculation (training set
1626 obs := make([][]bool, (maxv+1)*2) // 2 slices (hom + het) for each variant#
1627 // one-hot output (all samples)
1628 outcols := make([][]int8, (maxv+1)*2)
1629 for i := range obs {
1630 obs[i] = make([]bool, cmd.trainingSetSize)
1631 outcols[i] = make([]int8, len(cmd.cgnames))
1633 for cgid, name := range cmd.cgnames {
1634 tsid := cmd.trainingSet[cgid]
1635 cgvars := cgs[name].Variants[tagoffset*2:]
1636 tv0, tv1 := remap[cgvars[0]], remap[cgvars[1]]
1637 for v := tileVariantID(1); v <= maxv; v++ {
1638 if tv0 == v && tv1 == v {
1640 obs[v*2][tsid] = true
1642 outcols[v*2][cgid] = 1
1643 } else if tv0 == v || tv1 == v {
1645 obs[v*2+1][tsid] = true
1647 outcols[v*2+1][cgid] = 1
1652 var xref []onehotXref
1654 for col := 2; col < len(obs); col++ {
1655 // col 0,1 correspond to tile variant 0, i.e.,
1656 // no-call; col 2,3 correspond to the most common
1657 // variant; so we (normally) start at col 4.
1658 if col < 4 && !cmd.includeVariant1 {
1662 maf = homhet2maf(obs[col : col+2])
1663 if maf < cmd.pvalueMinFrequency {
1664 // Skip both columns (hom and het) if
1665 // allele frequency is below threshold
1669 if maf > cmd.maxFrequency {
1670 // Skip both columns if allele
1671 // frequency is above threshold
1676 atomic.AddInt64(&cmd.pvalueCallCount, 1)
1677 p := cmd.pvalue(obs[col])
1678 if cmd.chi2PValue < 1 && !(p < cmd.chi2PValue) {
1681 onehot = append(onehot, outcols[col])
1682 xref = append(xref, onehotXref{
1684 variant: tileVariantID(col >> 1),
1693 func homhet2maf(onehot [][]bool) float64 {
1694 if len(onehot[0]) == 0 {
1698 for i := range onehot[0] {
1702 } else if onehot[1][i] {
1707 return float64(n) / float64(len(onehot[0])*2)
1710 // convert a []onehotXref with length N to a numpy-style []int32
1711 // matrix with N columns, one row per field of onehotXref struct.
1713 // Hom/het row contains hom=0, het=1.
1715 // P-value row contains 1000000x actual p-value.
1716 func onehotXref2int32(xrefs []onehotXref) []int32 {
1718 xdata := make([]int32, 6*xcols)
1719 for i, xref := range xrefs {
1720 xdata[i] = int32(xref.tag)
1721 xdata[xcols+i] = int32(xref.variant)
1723 xdata[xcols*2+i] = 1
1725 xdata[xcols*3+i] = int32(xref.pvalue * 1000000)
1726 xdata[xcols*4+i] = int32(-math.Log10(xref.pvalue) * 1000000)
1727 xdata[xcols*5+i] = int32(xref.maf * 1000000)
1732 // transpose onehot data from in[col][row] to numpy-style
1733 // out[row*cols+col].
1734 func onehotcols2int8(in [][]int8) []int8 {
1740 out := make([]int8, rows*cols)
1741 for row := 0; row < rows; row++ {
1742 outrow := out[row*cols:]
1743 for col, incol := range in {
1744 outrow[col] = incol[row]
1750 // Return [2][]uint32{rowIndices, colIndices} indicating which
1751 // elements of matrixT[c][r] have non-zero values.
1752 func onehotChunk2Indirect(matrixT [][]int8) [2][]uint32 {
1754 for c, col := range matrixT {
1755 for r, val := range col {
1757 nz[0] = append(nz[0], uint32(r))
1758 nz[1] = append(nz[1], uint32(c))